I was originally planning on implementing tools (#898) as part of llm.Conversation but I've been having second thoughts about that.

There's a lot to be said for allowing prompts to reply to other responses, and having those form a chain.

Two big advantages:

• The in-memory (no SQLite database at all) story for the Python client library is simplified. Users don't have to remember to instantiate a llm.Conversation if they are going to send prompts that follow up on other prompts - particularly useful for implementing tools.
• Means I can support branching conversations. Maybe you start a conversation with a huge document (which gets into the prompt cache) and then ask multiple questions of that document as separate branches of replies to that initial prompt. Also good for a UI that lets power users really take control of the context in their ongoing conversations.